.. include:: ../links.rst


.. _dti_model:

##############################
Diffusion tensor imaging (DTI)
##############################


***************
DTI in QSIRecon
***************

DTI reconstruction is supported in QSIRecon using the
DSI Studio package (see :func:`~qsirecon.workflows.recon.dsi_studio.init_dsi_studio_recon_wf`)
or the TORTOISE package (see :func:`~qsirecon.workflows.recon.tortoise.init_tortoise_estimator_wf`).

The DIPY approach is accessible in a reconstruction specification by using a node with
``action: DTI_reconstruction`` and ``software: Dipy``.
Also see :class:`qsirecon.interfaces.dipy.TensorReconstruction` (primary DTI node),

The TORTOISE approach is accessible in a reconstruction specification by using a node with
``action: estimate``, ``software: TORTOISE``, and the ``parameters: estimate_tensor`` subdictionary.
Also see :class:`qsirecon.interfaces.tortoise.EstimateTensor`.


***********************
DTI Foundational Papers
***********************

**DTI Concept Introduction**: Basser et al. introduced DTI as a new modeling
framework that computes a Gaussian diffusion tensor per voxel, yielding eigenvalues
and eigenvectors that describe 3D water diffusion :cite:p:`basser1994`.
This enabled mapping of water diffusion orientations in tissue and orientation-independent
scalar measures like trace (sum of eigenvalues, related to Mean Diffusivity, MD).
It established the idea that tensors encode more microstructural information
than single-direction diffusivities.

**Anisotropy quantification**: Pierpaoli and Basser defined fractional anisotropy
(FA), a normalized 0-1 index quantifying how anisotropic (directionally dependent)
diffusion is :cite:p:`pierpaolibasser1996`.
They demonstrated that earlier methods (diffusion measured in just a few directions)
underestimated anisotropy, and introduced rotationally invariant metrics (FA, relative
anisotropy, etc.). This work also
cautioned that noise can bias anisotropy measures requiring eigenvalue ordering.

**First Human DTI Maps**: Pierpaoli et al. acquired the first DTI scans of the human
brain, mapping principal diffusion directions and magnitudes
:cite:p:`pierpaoli1996`. They observed that water diffuses ~3 times faster along
axonal fibers than perpendicular to them in highly coherent tracts (e.g. corpus
callosum).
They also introduced Trace (D = 3×MD) as
an orientation-invariant measure of overall diffusivity, which was roughly
uniform in normal brain except higher in cortical gray matter due to its
higher water content.

**Eigenvalue-Derived Metrics**: By the early 2000s, researchers began interpreting
individual tensor eigenvalues. Song et al. first showed that
axial diffusivity (AD, diffusion along the primary eigenvector, λ1) and
radial diffusivity (RD, diffusion perpendicular to axons, mean of λ2 & λ3)
can provide pathologically specific insights :cite:p:`song2002`.
In a mouse model of demyelination, RD increased with myelin loss while AD stayed
constant (since axons remained intact). This seminal finding
established that increases in RD selectively indicate myelin degeneration, whereas
decreases in AD are more tied to axonal injury - a distinction that has since
informed numerous neuroimaging studies of white matter diseases.

*******************************
DTI Studies Across The Lifespan
*******************************

The tensor model has been used extensively in the human brain
:cite:p:`pierpaoli1996` including developmental neuroscience.

**Neonatal Diffusivity**:  Neil et al. reported neonatal
mean diffusivity values 1.5-2 times higher than in adults, with very low
white-matter anisotropy :cite:p:`neil1998`. This reflects abundant free water and
unmyelinated fibers at birth. Diffusivity drops and anisotropy rises steeply
in the first postnatal months as the brain matures. Water compartmentalizes and
myelination progresses :cite:p:`ouyang2019delineation`.

**Childhood to Adolescence**: White matter development continues through
childhood and the teen years. Longitudinal data showed steady FA increases
and MD decreases in virtually all major tracts from age 5
into the 20s :cite:p:`lebel2011`. Not all tracts mature simultaneously:
early-developing motor/sensory pathways (e.g., internal capsule) reach adult-like
FA by late adolescence, whereas association tracts in frontal and temporal lobes
keep increasing in FA (and decreasing in RD) into the third decade. This prolonged
maturation of frontal circuitry aligns with functional development of executive
and cognitive abilities in late adolescence.

**Whole Lifespan Trajectories**: Cross-sectional analyses across the lifespan find
that FA follows an "inverted U" trajectory: increasing from childhood to a peak
in the 20s-30s, then declining with older age. In a sample of 430 subjects
aged 8-85 :cite:p:`westlye2010`, fractional anisotropy plateaus by the early 30s
and slowly falls thereafter, while mean and radial diffusivities do the reverse
(minimal in young adults, then rising in aging). Interestingly, this large study
found no simple "last-in-first-out" pattern although late-maturing frontal tracts
often showed pronounced aging changes, all regions eventually exhibited
microstructural decline, indicating a widespread but heterogeneous aging effect
rather than one specific sequence.

**Regional Patterns in Aging**: Many DTI studies of aging report that anterior white
matter tracts (which myelinate last) are more vulnerable to aging. Salat et al.
found significantly lower FA in older adults (mean age ~67) compared to young
(mean ~24), especially in the frontal lobes and corpus callosum :cite:p:`salat2005`.
In contrast, posterior tracts like the splenium of the callosum or occipital white
matter showed smaller FA differences. Such findings support that age-related myelin
degeneration and fiber loss are often greatest in late-developing, more complex pathways
(though subsequent research has refined this view with more nuanced patterns).

**Microstructural Changes with Aging**: DTI metrics suggest that aging involves
lower FA and increased water mobility
(higher MD/apparent diffusion coefficient, ADC) :cite:p:`westlye2010`.
In older adults, increased radial diffusivity is commonly observed, consistent
with demyelination or degraded myelin packing, while axial diffusivity may also
eventually decrease if axonal loss occurs. Longitudinal studies in elderly cohorts
(e.g. over 60) have confirmed ongoing within-person FA declines annually.
These DTI changes correlate with cognitive slowing and
executive function decline in many studies, highlighting DTI's value in tracking
brain aging and its cognitive consequences.

***************************************
DTI Methodological Warnings and Caveats
***************************************

**Single Tensor Limitations (Crossing Fibers)**: The basic DTI model assumes one
dominant fiber orientation per voxel - an assumption often violated in the brain.
In regions with crossing, kissing, or branching fibers, the tensor model yields
an average that can underestimate anisotropy and obscure fiber directions. For
instance, a voxel containing two crossing tracts will show an artificially low
FA (appearing "isotropic") even if each tract is highly anisotropic, and principal
diffusion direction that is an average of the true tract directions.
This issue
can lead to misinterpretation of reduced FA: it might reflect complex fiber
geometry rather than neural degeneration. Advanced high-angular-resolution methods
(e.g., multi-tensor or Q-ball imaging) are recommended when crossing fibers are
prevalent, or one should interpret DTI metrics in such regions with caution
:cite:p:`wheelerkingshott2009`.

**Axial vs. Radial Diffusivity Interpretations**: While increases in radial
diffusivity and decreases in axial diffusivity have been linked to demyelination
and axonal injury respectively, one must be careful not to over-interpret these
metrics in isolation. Wheeler-Kingshott and Cercignani showed that in voxels with
multiple fiber orientations, a change in one eigenvalue
can induce a "fictitious" change in another :cite:p:`wheelerkingshott2009`.
For example, crossing fibers can make AD appear reduced even without axonal damage.
Similarly, heavy pathology can alter the principal eigenvector direction,
invalidating the simple mapping of λ1 to one fiber population.
*Bottom line*: AD and RD are informative only in contexts where a single fiber
population dominates the voxel; otherwise, observed changes might result
from geometry or partial volume effects rather than specific
histopathology.

**Partial Volume and Free Water Contamination**: DTI metrics can be skewed by
mixing of tissue with free water (cerebrospinal fluid or edema). A small amount
of free water in a voxel drastically lowers FA and raises diffusivity, since free
water diffusion is fast and isotropic. This can mask true tissue changes - for
example, a remyelinating lesion adjacent to CSF might still show low FA due to
CSF contamination. Methods like the free-water elimination model address this by
fitting a two-component model :cite:p:`pierpaoli2004`,
effectively stripping out the isotropic diffusion component and revealing the
"true" tissue tensor. Researchers should be mindful of partial voluming, especially
in periventricular areas, and consider correction strategies or region-of-interest
approaches to avoid artifactual findings.

**Noise, Motion, and Artifacts**: DTI outcomes are sensitive to data quality.
Thermal noise in diffusion-weighted images leads to bias (e.g., a noise floor
artificially boosts low ADC values), and insufficient signal-to-noise can make
FA appear higher in low-FA regions (background noise imposes a floor)
:cite:p:`jones2010`. Head motion is another major concern: even subtle
motion can cause directional-dependent blurring or signal drop-out, which may
mimic or mask true diffusion anisotropy. For example, uncorrected motion can
spuriously increase FA in gray matter or produce group differences unrelated to
biology (as noted in studies of populations like children or patients who move
more) :cite:p:`yendiki2014`. Best practices include using motion correction
algorithms, excluding data with excessive motion, and using robust fitting methods
(e.g., iteratively reweighted least squares) that down-weight outliers caused by
artifacts.

**Acquisition and Analysis Choices**: The choice of diffusion gradient directions
and b-value can influence DTI metrics. A minimal 6-direction tensor encoding is
insufficient for reliable quantitative work - more directions (20-30+) are
recommended to stabilize FA/MD measures and reduce variability :cite:p:`jones2004-hs`.
Similarly, moderate b-values (~1000 s/mm²) are typically chosen to balance SNR and
sensitivity; very high b-values can introduce bias in tensor-fitting,
due to higher sensitivity to non-Gaussian diffusion --
and require other models, such as DKI. During analysis, image alignment (registration)
and smoothing can also introduce caveats: misregistration across subjects can
blur tract-specific values, and heavy smoothing can artificially decrease FA
in partial volume voxels. The key caveat is that DTI analyses involve many
processing steps, each of which must be done carefully - otherwise, errors can
propagate and lead to incorrect conclusions. Community guidelines and
detailed "pitfall" checklists (e.g., :cite:p:`jones2010`) are valuable
resources to ensure methodological rigor in DTI studies.

**********
References
**********

.. bibliography::
   :style: unsrt
   :filter: docname in docnames